RFR(XS): 8171092: C1's Math.fma() intrinsic doesn't correctly process its inputs
vladimir.kozlov at oracle.com
Mon Dec 12 17:23:12 UTC 2016
Looks good. When jdk9/hs is open for pushes I will sponsor it.
On 12/12/16 6:32 AM, Roland Westrelin wrote:
> 3rd input of the fma intrinsic is processed as a temp but is actually an
> input. The test case triggers the problem on aarch64 (with the patch
> from 8162338). I couldn't get a failure on x86 because the input is
> overwritten by the fma instruction: a copy of the input is made to a new
> register right before the fma instruction and the input is live until
> right before the fma instruction. I'm far from certain the bug cannot
> show up on x86 with some other code pattern.
More information about the hotspot-compiler-dev